Elevation Change prediction model · Basketball
The Bot
Our own numbers, per possession. Every Division I team rated, every game picked before tip, and every miss kept on the record.
Win chances come from the Bot’s projected margin and how far real games stray from it. Bot line = the Bot’s own spread. Pace = projected possessions. Picks lock at tip.
Net rating = points per 100 possessions better than an average Division I team on a neutral floor. Offense is points scored above average, defense is points prevented. Tempo = possessions per 40 minutes against an average opponent. Back = share of last season’s minutes returning (shown once a team’s roster is updated).
From 1,000 simulated seasons, rebuilt every Monday with the bracket. Title = regular-season league title (ties shared). Tourney = winning the league tournament and its automatic bid. NCAA = making the 76-team field either way; the small line splits it into automatic and at-large. Seed = projected seed if the season played out as the Bot expects.
The Bot is Elevation Change’s own basketball model. It never reads betting lines, KenPom, the NET or anyone’s rankings to make a pick. Lines are used only to grade it.
1. Ratings per possession
Basketball is measured per possession, so a fast team and a slow team can be compared fairly. Every game gives each team an offensive and defensive efficiency: points scored and allowed per 100 possessions. The Bot solves for the offense, defense and tempo of every Division I team at once, adjusting each game for who the opponent was and where it was played. Blowout margins are capped so running up the score doesn’t pay.
2. Where teams start
Before a game is played, every team starts from last season’s ratings, pulled toward its conference and adjusted for returning minutes and production (who’s back, and how good they were), incoming transfers (what each one produced at his old school, weighted by how good that school was) and ranked high school recruits (On3’s Industry ranking: a top-10 recruit is worth about 2.3 points per 100 possessions, a top-25 about 1.6, a top-100 about 0.8). Those weights are fit on thirteen seasons of what actually happened. Preseason exhibitions against Division I teams also nudge the start: each counts as a full game, graded against what the Bot expected, then fades out once the season starts and stops counting a quarter of the way through the regular season (early December). They come from the schools’ own schedule pages, since ESPN doesn’t carry exhibitions, so there’s no history to fit that weight; teams do experiment in exhibitions, which is why they fade out quickly. The preseason start counts as about four games, so by mid-December the ratings are almost all this season. The preseason start only drives the Bot’s picks; the Elevation Index drops it once teams have two games in and ranks on résumé and this season’s stats.
3. The game model
4. Keeping it honest
- Locked at tip. Each pick freezes at the last update before the game starts. The record is graded against those locked picks, never a pick made after the fact.
- Tested on games it never saw. The model was fit on seasons through 2020-21, then asked to predict every game since, one day at a time, using only what had happened before that morning.
- Vegas only for grading. Pregame lines are saved as games come up so the Bot can be measured against them. They never feed a pick.
- Data: schedules, scores and box scores from ESPN via the public hoopR archive, plus ESPN’s scoreboard for same-night finals. Updates every morning and through game nights.